Fourth International Conference on Geotechnique, Construction Materials and

Environment, Brisbane, Australia, Nov. 19-21, 2014, ISBN: 978-4-9905958-3-8 C3051

Practical Application of Probabilistic Methods in Geotechnical Engineering

Stephen Buttling, BSc(Eng), ACGI, PhD, CEng, FICE, FIEAust, CPEng

GHD Pty Ltd, Brisbane, Australia

ABSTRACT

Despite the fact that probabilistic methods in engineering have been around for about 50 years, and their
undoubted relevance to geotechnical engineering as a result of data uncertainty, they have not been widely
adopted in everyday geotechnical design practice. There have been many publications, both books and technical
papers, by eminent academic authors, but the knowledge of statistics required to make good use of the methods
is held by few. In recent years, readily available commercial software has made it possible for probabilistic
analyses to be carried out extremely easily, perhaps too easily, as the basic statistical knowledge is still required.
Examples of misunderstandings of methods and misinterpretations of results will be shown, together with
suggested ways in which these can be overcome and more appropriate use of the very powerful methods made.

Keywords: Probabilistic analysis, slope stability, probability of failure, computer software

INTRODUCTION acknowledge that engineering in general, and

geotechnical engineering in particular, is concerned
Probabilistic methods have been around in civil with parameters which have a high level of
engineering, and in geotechnical engineering, for uncertainty. Ang & Tang [3] note that “the role of
quite some time, but yet they are not widely used probability and statistics is quite pervasive in
and accepted in practice. If we look carefully at the engineering.” This applies not only to soil properties,
work which has been done, it tends to be centred in but also to rainfall and hence run-off quantities,
academia, where the necessary understanding and wind forces on buildings, earthquake magnitudes,
knowledge can be expected to be available. On the traffic volumes, and many more examples. In some
other hand, the day to day practice may be left to of these the role of statistics is acknowledged and
engineers with a far lower level of skill, and this can, incorporated, especially in the case of rainfall, but in
and indeed has, led to significant misunderstandings others we rely more on a factor of safety approach,
and misinterpretations of results. believing that this will provide adequate protection,
without needing to know how much. It is worth
HISTORICAL PERSPECTIVE noting that the quotation from Ang & Tang [3] refers
to the role of statistics and not to the use.
Although they were specifically made with The chart in Figure 1 was first published by
regard to the application of Bayesian Methods to Christian et al [6], and has since been repeated by
updated parameter values based on some prior many others, including Nadim et al [7] and Pollock
knowledge plus some site specific testing, the et al [8].
comments in Simpson & Driscoll [1] are believed to
be relevant and have a wider implication:
The necessary skills in statistics are held by
“very few designers who have committed their time
to training and experience in geotechnical
engineering. Attempts by statisticians to tackle
geotechnical design have often ended in ridicule,
and it is very difficult for one person to have a
sufficient grasp of both disciplines that he can
combine them sensibly.”
Nevertheless we have excellent text books, such
as two editions of Ang & Tang in 1975 [2] and
2007 [3], Duncan & Wright [4], and Phoon [5]
amongst others, but their influence does not seem to
have made its way into everyday engineering
practice, and in particular into the assessment of Fig. 1 Effect of data uncertainty on probability of
slope stability. failure for two mean Factors of Safety
This is despite the fact that most of us

1
 GEOMATE- Brisbane, Nov. 19-21, 2014

It indicates the type of faith which practitioners standard deviation below the mean, then this will
have in the traditional use of factors of safety, give a more conservative answer. For simplicity of
derived from deterministic analyses, without demonstrating the concept it is assumed that the
considering data uncertainty. Because statistical distribution of factor of safety will be similar to the
analyses are frequently associated with large distribution of strength, and therefore use of a
volumes of data, there is a very understandable characteristic strength half a standard deviation
concept that they can only be used in geotechnical below the mean strength will be equivalent to a
engineering when we have large volumes of data, Factor of Safety half a standard deviation below the
and that is rare. However there is no doubt that we mean Factor of Safety, then we can find the
are dealing with data uncertainty, and statistical distribution of Factor of Safety which will give a
methods also have wide application when we have characteristic Factor of Safety of 1.5 or 1.2 as in
very limited data. Duncan [9] has suggested the 6 Figure 1. This is shown in Figure 2.
sigma method amongst others, when there is little or It is also noted that it has been recommended that
no data, and this involves selecting the smallest slope stability analyses should be carried out using
possible value and the largest possible value, and at least two computer programs, because of the
assuming that they are separated by six times the complexities of the different algorithms within each
standard deviation. Schneider [10] suggested program. In geotechnical engineering consultancy
selecting the lowest (a), highest (c) and most likely practice this is unlikely to occur as a result of
(b) values, and then defining the mean as: commercial pressures both on time and licences.
௔ାସ௕ା௖
ߤൌ (1)
଺ BASIS OF METHODS
and the standard deviation as:
௖ି௔
ߪൌ ଵଶ
(2) In general the probabilistic methods involve
determination of a distribution of values of a
Using either of these methods to consider the likely
parameter, rather than a unique value as is used in
range of values of a parameter within a probabilistic
deterministic analyses. Today we may also use
analysis will probably produce a more meaningful
statistical methods to determine our unique value as
output than picking a single value in a deterministic
a characteristic value, to represent the data set, or the
analysis.
combination of a new data set with some previous
This is not to say, however, that probabilistic
knowledge. It has been found by various authors, e.g.
analyses should replace deterministic analyses
Ang & Tang [3], and Bond & Harris [11], that soil
entirely. There is clearly a need for both, but
strength parameters, being “physical properties
probabilistic analyses are almost certainly not be
based on a large number of individual, random
given appropriate weight at the present time because
effects”, will often be normally distributed about the
of misunderstandings over their applications and
mean value. However, great care is needed not to
benefits.
overuse this familiar and convenient model. Since
the spread of the distribution needs to be defined as
well as the mean value, the standard deviation is
commonly used as a measure of spread. This is also
often expressed as the Coefficient of Variation, or
CoV, defined as:

‫ ܸ݋ܥ‬ൌ ߪൗߤ (3)

where σ is the standard deviation and μ is the mean

of the sample. Since it is known that 99.7% of all
values in a normal distribution lie within the range
of μ ± 3σ, it follows that if CoV > 33% then some
values will be negative. Data published by Duncan
[9], Nadim et al [7] and Bond & Harris [11] has
Fig. 2 Modified distribution with characteristic FoS shown that CoV values for parameters such as
= 1.5 and 1.2 apparent cohesion, c’, undrained shear strength, su,
and coefficient of volume compressibility, mv, can
In fact the curves in Figure 1 are considered to be range up to 40, 50 or even 70%, this would produce
slightly misleading, since it is considered that it is negative values if used with a normal distribution.
not normal practice in geotechnical engineering to Since for these and most engineering properties
use mean strength for design. If we consider that a negative values are not permissible, the normal
characteristic strength should be used, and that the distribution is unlikely to be appropriate if CoV >
characteristic strength may be defined as half a 30%. This has also been discussed by Look [12].

2
 GEOMATE- Brisbane, Nov. 19-21, 2014

On a project in Queensland, cement was being of twelve profiles are shown in Figure 3, which
mixed with a bi-product of the alumina industry, illustrates the scatter in the results. It is also apparent
known as red mud, in order to strengthen it to allow that many strength values are over 300 kPa, with
an upstream raise of the tailings dam. Trials were some up to 700, while negative values are not
carried out, and in situ tests were used to measure possible. As a result the data does not fit a normal
the undrained shear strength of the treated material. distribution.
Figure 4 shows the data for each profile plotted
as a cumulative distribution function, and it can be
seen that some follow a smooth trend, while others
involve significant changes in gradient identify gaps
in the strength distribution.
What is remarkable is that, when all the data are
combined, making over 2,500 data points, the
cumulative distribution function has a very close fit
with a log-normal distribution curve, as shown in
Figure 5.

Fig. 5 Cumulative distribution function for 12

profiles combined

This would suggest strongly that a log-normal

distribution is a good representation of the strength
of this modified material. In other recent examples, a
Fig. 3 Wing Cone Penetrometer profiles
beta distribution was found to give a good
representation of the probability distribution of some
shear wave velocity results, seen in Figure 6, a log-
normal distribution was found to work for some
Lugeon test results seen in Figure 7, and an inverse
Gaussian distribution gave the best fit for some
Point Load Index test results seen in Figure 8.

Fig. 4 Cumulative Distribution Function plots for 12

profiles of WCP data

The tests were carried out using a device called a

Wing Cone Penetrometer, developed in Sweden as a Fig. 6 A gamma distribution fitted through some P-
variation on a Cone Penetrometer Test to smooth wave velocity data
over minor spatial variations in strength. The results

3
 GEOMATE- Brisbane, Nov. 19-21, 2014

function for the data in Figure 5, as well as beta and

gamma distributions which would be unlikely to be
distinguishable when applied to actual data sets in
real geotechnical engineering problems.

SPECIFIC METHODS

USE OF PROPRIETARY SOFTWARE

It is significant that, in the field of slope stability,
two of the programs believed to be most widely in
industry both offer probabilistic analysis as an
Fig. 7 A log-normal distribution fitted through some option. This has made it easy for geotechnical
Lugeon test data engineers to carry out such analyses but,
unfortunately, it has not ensured that they have the
necessary skills. This is one of the common features
of modern proprietary software. Many years ago
such software was written and presented in complex
code or programming language, and only people
skilled in its use could understand and operate it.
Now everything is windows-based, and user-
friendly, and it has become increasingly difficult to
ensure that software is only available to those with
the requisite skills to operate it. This is especially
true with the younger generation of computer literate
engineers, who spurn hand calculations, and readily
Fig. 8 An inverse Gaussian distribution fitted assimilate such things as finite element modeling
through some Point Load Index data without needing a grounding in the theory of
stiffness matrices or constitutive modeling.
Returning to the subject of slope stability, it is
very easy to click on the buttons in a program such
as SLOPE/W to carry out a probabilistic analysis,
without understanding how the analysis is carried
out, or what the results mean. Examples have been
published in which the authors have carried out
probabilistic analyses, without understanding that
they have not modeled spatial distribution of
strength. By choosing a very simple geotechnical
model, in order not to have distractions from the
main point they were trying to make, the Monte
Carlo method was only able to randomly select
strength parameters, in this case c’ and φ’, and then
apply them to the whole slope since there was a
Fig. 9 Comparison of some skewed distributions single soil layer. The resulting probability of failure,
based on the proportion of critical failure surfaces
It is clearly essential to have a reasonable which returned a factor of safety of < 1, was
estimate of the probability distribution of the therefore the probability that the whole slope had
required parameters, in order to be able to select uniform properties which did not provide enough
appropriate values to use in a probabilistic analysis. resistance.
Nevertheless it is noted that several have a similar This helps to emphasise the difference between
shape, such as the beta, gamma, and log-normal probabilistic and deterministic analyses. In the
distributions, and it is unlikely that a significantly former, we choose suitably conservative strength
different outcome will result from their use. It is parameters and apply them to the whole slope. But
worth noting that the individual real data profiles in in the latter, in order for the probabilistic analysis
Figure 4 did not follow an idealized log-normal and the Monte Carlo simulation to have any
profile closely, while the combination of all the data meaning, there must be some realism to the random
into one set produced the very close agreement with sampling, and applying one set of strength
the log-normal curve shown in Figure 5. Figure 9 parameters uniformly to the whole layer is not it.
shows the log-normal probability distribution Within the context of a probabilistic slope stability

4
 GEOMATE- Brisbane, Nov. 19-21, 2014

analysis, we need the random sampling to be considered. This means that the Monte Carlo
modeling the spatial variability of strength, so we method, applied through 100,000 simulations, has
need there to be variations of strength within each randomly picked values for c’ and φ’, based on the
layer. mean and standard deviation supplied as determined
In order to investigate this, the simple slope Table 1 Probabilistic results from basic analysis
model shown in Figure 10 and analysed by Pollock
et al [8] has been re-examined. First their results Mean F of S  1.1309 
were regenerated, but it was found that the use of the Reliability Index  0.707 
optimization feature of SLOPE/W combined with P (Failure) (%)  26.925 
Standard Dev.  0.185 
Min F of S  0.69976 
Max F of S  1.5791 
# of Trials  100000 

by Pollock et al [8], and then applied them to the

whole slope. While it is understood that this is a
special simple case chosen to illustrate a point with
regard to the use of probabilistic analysis, it is also
important to realize that such a distribution of
strength will never occur in practice, since the
random spatial distribution has not been modelled.
Again this is an extreme case which has made the
inappropriateness of this model clear because it is
Fig. 10 Basic 15° slope analysed after Pollock et al applied to the whole slope, but extreme care would
[8] also be needed when applying such a model to a
multilayered slope, to ensure that the spatial
variation of strength has been appropriately
modelled.
The parameters which affect this are the auto-
correlation distance, which relates to the distance
over which properties, such as strength, are found to
remain constant such that they can be modelled by a
single strength selection in a Monte Carlo process.
Using the same parameter over the whole area
modelled is equivalent to using an auto-correlation
distance = ∞. El-Ramly et al [13] reported that, for
the lacustrine clays in the James Bay area of
Quebec, Canada, auto-correlation distances of about
Fig. 11 Results of Morgenstern-Price analysis with 30 m were found to be appropriate. This has also
optimization been advised for quick clays in Norway. However,
for older soils which may have been heavily over-
the Bishop Method of Slices gave distorted results, consolidated and fissured, and for tropically
even though they matched the output given in the weathered residual soils, very much shorter auto-
paper. Instead the Morgenstern-Price method was correlation distances may be appropriate. Yu and
selected, and this gave the results shown in Figure Mostyn [14] suggested auto-correlation distances of
11 and in Table 1. 1 to 2 m.
A “standard” SLOPE/W analysis from Geostudio In order to investigate the effect of this on the
2007 was used, except that the optimization function slope in Figure 10, the geometry was modified to
was turned on as it had been by Pollock et al [8]. have a mosaic of squares of soil of 2 m side, as
The results were equivalent, except that they shown in Figure 12.
appeared to have used the Bishop Method of Slices, The soil model was cloned such that all squares
which gives a higher Factor of Safety when used were notionally different soils, but with identical
with the optimization function than either the properties. This allowed the Monte Carlo simulation
Morgenstern-Price method with optimization of the to allocate different strengths randomly to each “soil
Bishop Method without. square” which, it may be argued, is a reasonable
The most striking thing about the results in Table approximation to the random spatial distribution of
1 is the Probability of Failure at nearly 27%. Since strength in nature. The effect on the results was
no temporal variation has been allowed for in this dramatic.
analysis, it is only spatial variation that has been

5
 GEOMATE- Brisbane, Nov. 19-21, 2014

Fig. 14 Effect on probability of failure of number of

Fig. 12 Slope with mosaic of 2 m squares trials for model with 2 m soil squares

The slip surface and the mean factor of safety landslides, is the product of likelihood and
were basically unaffected after 100,000 trials as seen consequence. This in itself can lead to
in Figure 13, but the probability of failure had complications, because the general population may
dropped by nearly three orders of magnitude, as refer to risk in the context of probability, i.e. “the
shown in Table 2. risk that something will happen”, or of
susceptibility, i.e. “this is at a greater risk of slipping
than that”. However, accepting that in landslide risk
management, risk means the product of likelihood
and consequence, in practice it seems to be easier to
determine the consequences associated with a
landslide of given dimensions, than it is to determine
the likelihood that it will happen. There are a
number of reasons for this, and obviously knowing
the cause of a landslide is very high on the list,
especially when we are referring to the cause of a
landslide which has not yet happened. It is also
generally accepted that there is a link between
rainfall and landsliding, but details around the link
Fig. 13 Output from probabilistic analysis with are sketchy. What is the period of extreme rainfall
mosaic of 2 m soil squares which triggers landslides? Is it 10 minutes, an hour,
24 hours, 72 hours, or more? Is it different for
Table 2 Probabilistic results from 2 m square different types of landslides, such as natural slopes,
analysis or fill slopes?
Mean F of S  1.1294  All of this makes the determination of an annual
Reliability Index  3.423  probability of extremely difficult, but nonetheless
P (Failure) (%)  0.031  necessary if we are to establish likelihood in order to
Standard Dev.  0.038  evaluate risk. It has already been noted that the
Min F of S  0.97796  readily available commercial software for slope
Max F of S  1.2813  stability will run a probabilistic analysis and produce
# of Trials  100000  an average factor of safety and also a probability of
failure. What has happened is that users have not
The effect on the probability of failure of the stopped to question what that probability of failure
number of Monte Carlo simulations was tested, and actually means.
it was found that there was a small drop in the It is easy to see that there is a connection
probability as the number of trials was increased between annual probability of failure, and
from 10,000 to 200,000, as seen in Figure 14. probability of failure during a design life, with the
latter probability increasing with increasing design
ANNUAL PROBABILITY OF FAILURE life. This leads to the equation:

Another problem in application which has ࡼࡰࡸ ൌ ૚ െ ሺ૚ െ ࡼࢇ ሻ࢟ (4)

become evident relates to a misunderstanding of
annual probability of failure. It is now commonly In which PDL is the probability of failure during the
accepted that the term “risk”, when applied to design life of y years, and Pa is the annual
probability of failure. This is illustrated in Figure 15.

6
 GEOMATE- Brisbane, Nov. 19-21, 2014

deviation = 6.9 kPa. With his 5 m high slope at 1(V)

to 2(H) and 5 m below the toe to the rigid boundary,
he found a deterministic factor of safety = 1.356. He
then argued that the selection of the critical surface
based on uniform parameters, followed by a Monte
Carlo simulation of spatial variability, is unreliable
since the spatial variability will inevitably affect the
selection of the critical surface. This was
demonstrated by analysis using his Random Field
model, which is capable of modeling spatial
variability. The effect was to identify a number of
critical slip circles with factors of safety ranging
between 0.834 and 0.983. However, an examination
Fig. 15 Variation of probability of failure with of the contours of undrained shear strength used in
design life. the different models suggests that an experienced
geotechnical engineer would not carry out a
In theory it would be possible, if the exact deterministic analysis with an undrained shear
probability of failure during a given design life were strength of 23 kPa, because of the existence of large
known, to calculate the annual probability of failure. zones of lower strength material in what would be
In practice this is exceedingly unlikely to happen, expected to be critical zones.
since circumstances dictate that it is more practical
to determine an annual probability of failure than it Cho [16] also examined the influence of cross-
is to determine a probability of failure during a correlation between c’ and φ’ for a homogeneous
design life. Unfortunately, what has happened on at slope in his Example 2. He found the relationship
least two occasions published recently, is that shown in Figure 16 between Coefficient of Variation
analysts have confused probability of failure, based and cross-correlation, using both a fixed critical
on spatial variability, with probability of failure surface and his Random Field method.
based on temporal variability, and have equated the
spatial probability of failure with probability of
failure during a design life. They have then tried to
calculate an annual probability of failure by
assuming a design life. This is clearly incorrect,
since it leads to the conclusion that annual
probability of failure decreases with increasing
design life.
What needs to be done is that a reasonable
geotechnical model is built, in accordance with the
previous section, which makes due allowance for
spatial variability. Superimposed on this we can then Fig. 16 Variation of Coefficient of Variation with
place different groundwater levels, with appropriate cross-correlation coefficient after Cho [14]
annual return periods. It will then be straightforward
to determine a probability of failure which can be
combined with the temporal probability to produce
an overall annual probability of failure. These
probabilities for each groundwater level can then be
combined as in Table C6 in AGS [15].

USE OF CUSTOM MADE SOFTWARE

An examination of the literature tends to suggest
that probabilistic analysis is still mainly carried out
by academics within universities or research
institutes, where the level of statistical understanding
is suitably high, and time is available to write
dedicated software, even if it makes use of readily Fig. 17 Variation of Coefficient of Variation with
available materials such as Excel from Microsoft. cross-correlation coefficient using SLOPE/W
Another example of this is the work of Cho [16].
In Example 1 in this study Cho [16] also looked Attempts to reproduce this relationship in
at a uniform homogeneous slope, with a mean SLOPE/W using the same slope and soil parameters,
undrained shear strength = 23 kPa and a standard and with a spatial correlation distance = 20 m, was

7
 GEOMATE- Brisbane, Nov. 19-21, 2014

not successful as there was an unexplained peak in civil & environmental engineering. 2nd Ed. New
the curve around a cross-correlation = 0, as seen in York, Wiley 2007.
Figure 17. [4] Duncan JM & Wright SG, Soil strength and
This helps to illustrate the dilemma faced by slope stability. New York, Wiley, 2005.
practicing engineers. There is a lack of good [5] Phoon KK, Reliability-based design in
understanding of the statistical processes involved in geotechnical engineering – computations and
these analyses, and of the specialized jargon used by applications London, CRC, 2008.
statisticians which is indecipherable to most of us, [6] Christian JT, Ladd CC & Baecher GB,
and there are now computer packages which appear Reliability and probability in stability analysis,
to make the processes simpler, but involve the Journal of Geotechnical Engineering, ASCE,
extreme danger of being able to produce answers 120, 12, 2180-2207, 1994.
without needing to understand the mathematical [7] Nadim F, Einstein H & Roberds W,
processes. In design offices all over the world, Probabilistic analysis for individual slopes in
geotechnical engineers do not have the luxury of s0il and rock, International Conference on
time to write custom software, or even adapt Landslide Risk Management, Vancouver, 63-98,
spreadsheets in ways which have been suggested by 2005.
some authors (El-Ramly et al [13]; Low BK et al [8] Pollock D, Hurley G & Haberfield C, Linking
[17]; Wang et al [18]) equilibrium analysis and landslide risk
assessment, Australian Geomechanics, 46, 2,
CONCLUSIONS 149-162, 2011.
1 Probabilistic analyses are very powerful tools [9] Bond A & Harris A, Decoding Eurocode 7,
in the correct hands, and have particular London, Taylor & Francis, 2008.
application in geotechnical engineering as a [10] Duncan JM, Factors of safety and reliability in
result of data uncertainty. geotechnical engineering, Journal of
2 The necessary statistical skills are held by Geotechnical and Geoenvironmental
very few people, and geotechnical engineers Engineering, ASCE, 126, 4 307-316, 2000,
in general lack the necessary training in [11] Look B, Appropriate distribution functions for
statistics. characteristic values (to be published) 2013.
3 Changes are needed in the education of [12] El-Ramly H, Morgenstern NR & Cruden DM,
geotechnical engineers, to make them more Probabilistic slope stability analysis for practice,
aware of engineering statistics, the tools Canadian Geotechnical Journal, 39, 665-683,
available, and how to make proper use of 2002.
them. [13] Yu YF & Mostyn GR, Random field modelling
4 Proprietary software now offers some for the effect of cross-correlation, Proc. 13th
probabilistic analysis capability, but it is not International Conference on Soil Mechanics and
well documented and explained, leading to Foundation Engineering, New Delhi, 4, 1389-
serious errors in analysis and reporting. 1392, 1993.
5 Although it has been recommended that slope [14] AGS, Commentary on Practice Note Guidelines
stability analysis is carried out using at least for Landslide Risk Management 2007,
two computer programs, in actual Australian Geomechanics, 42, 1, 115-158, 2007.
geotechnical engineering practice this is [15] Cho SE, Probabilistic assessment of slope
unlikely to occur due to commercial stability that considers the spatial variability of
pressures. soil properties, Journal of Geotechnical and
Geoenvironmental Engineering, ASCE, 136, 7,
REFERENCES 975-984, 2010.
[16] Low BK, Practical probabilistic slope stability
[1] Simpson B & Driscoll R, “Eurocode 7 – a analysis, Proc. 12th Panamerican Conference on
commentary”, BRE Report BR 344, Soil Mechanics and Geotechnical Engineering,
BRE, London, 1998. MIT, Cambridge, Massachusetts, 2, 2777-2784,
[2] Ang AH-S & Tang WH, Probability Concepts 2003.
in Engineering Planning and Design, Vol I, [17] Wang Y, Cao Z and Au SK, Practical reliability
Basic Principles. New York: Wiley, 1975. analysis of slope stability by advanced Monte
[3] Ang AH-S & Tang WH, Probability Concepts Carlo simulations in a spreadsheet, Canadian
in Engineering – Emphasis on applications in Geotechnical Journal, 48, 162-172, 2011.